Use Claude extended thinking together with tool use, preserving thinking and redacted_thinking blocks correctly across tool_result turns

Documented steps

1. Enable manual thinking with thinking:{"type":"enabled","budget_tokens":N}. The minimum accepted budget is 1024, and N must be less than max_tokens because thinking tokens count toward the turn's max_tokens.
2. For interleaved thinking between tool calls, add anthropic-beta: interleaved-thinking-2025-05-14. Under that beta budget_tokens may exceed max_tokens because the budget spans all thinking blocks in the turn.
3. Check model support before relying on any of this: interleaved thinking is unsupported on some models and deprecated on others, and the newest model generations reject manual thinking:{type:'enabled'} with a 400. Read the current model table rather than assuming.
4. Do not send non-default temperature, top_p, or top_k on the newest models — they return 400 regardless of thinking. On older models temperature and top_k conflict with thinking, and top_p is only allowed between 0.95 and 1.
5. With manual thinking, only tool_choice {"type":"auto"} or {"type":"none"} is legal. Forced tool choice ({"type":"any"} or a named tool) errors.
6. When you send the tool_result back, return the assistant's thinking blocks unmodified in the same order. This is required within a tool-use turn.
7. Return redacted_thinking blocks ({"type":"redacted_thinking","data":"..."}) unchanged too. Filtering content on block.type == 'thinking' silently drops them and breaks the protocol.
8. Never edit, reorder, or partially drop the consecutive thinking blocks in the latest assistant message — the API rejects modified blocks with a 400.
9. Each thinking block carries a signature the API uses to verify provenance; when streaming, it arrives as a signature_delta just before that block's content_block_stop.
10. Hold the thinking configuration constant for the life of a cached conversation: changing mode, budget_tokens, or effort invalidates the prompt cache prefix.
11. For very large thinking budgets, prefer the Message Batches API to avoid long-connection timeouts.
12. Official docs: https://platform.claude.com/docs/en/build-with-claude/extended-thinking